Runaway Reaction: When Benign Skills Compose into Malicious Behavior
Paper finds vetted benign agent skills can compose into malicious behavior and introduces a 4,000-skill benchmark.
The paper argues that marketplace agent skills are usually vetted alone, so risks created only by composition are missed. The authors report that combining individually benign skills can induce malicious behavior even when every skill passes standalone checks. They present CRIME, which casts a target behavior into complementary requirements, selects public skills, and evaluates skill pairs in a sandbox, with iterative steering when direct composition fails. They also construct a benchmark of 4,000 public skills across eight cybersecurity behaviors.
- Isolated skill vetting misses risks that appear only after composition.
- Directly combining vetted benign skills can already produce malicious behavior.
- CRIME uses plot casting, steering, and sandboxed skill-pair execution.
- Authors built a 4,000-skill benchmark covering eight cybersecurity behaviors.
From the original article
1 min read at arxiv.orgAgent skills package task-specific knowledge and procedures that can be composed to support complex agent tasks, while public marketplaces provide a growing pool of reusable skills. Existing security vetting, however, largely evaluates skills in isolation, leaving composition-induced risks underexplored.
By Zunlong Zhou · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.